How Relevant Is Training in a Workforce That Is AI Agents and Some Humans?
A static ledger of skills dissolving while the workforce changes shape.
Exhibit A: The Tooling Notes
Claude Docs: use batch to birth a doc and update for edits. Use search or view reads, and guard whole-block replace or delete with ifHash and ifRev. Open each new doc with the Artifact open action.
For diagrams, read the diagram guide first. They are SVG modules, and per his preference they have no clickable nodes.
Web research: use WebFetch on primary pages, and cite only pages actually opened, with an as-of date.
Skill, Capability, and the Rate of Change
The text in Exhibit A represents a skill. I will come back to why that matters.
First it is important to state that, in my reading, training is not the answer to disruption caused by AI. Replacement is already underway, and the workforce that results is AI agents and some humans, not humans and AI. At the companies moving fastest, the answer is replacement, and they are saying so in their filings and memos. The rest of the industry is moving slowly, layering AI onto structures built before AI came along (skills, competency, etc.).
I learned what a skill is in professional management college, where I was trained to manage complex operations, to hire and train people, to balance human growth and balance sheets. The day ran 9 am to 1 pm in practical classes, 2 to 5 pm learning theory, and then every evening I worked in the industry. I learned, applied it in the workshop, then did it for real, every day. I came out of college a manager, like most of my batch-mates. A skill, I learned, is an integration of run-time context, real-time information, behaviour, deep knowledge and decision making, and it is never fixed because the context never is. I then spent 12 years in operational roles, hiring, training, coaching and firing people on the shop floor, writing SOPs, assessing skill and making decisions every minute with almost no margin for error.
Then I moved into learning technology. In my opinion, the eLearning and L&D industry's approaches are built to deliver at scale to address human workforce development problems that arose due to the ascendance of information technology and related new services that became integral to industries. The frameworks that evolved addressed skills as static for long periods and covered mainly the cognitive component, because that is the part that can be delivered quickly at scale - which was the core problem then. "Skill-first" names what the industry needs, but it is deployed as a static, cognitive-only framework. The methods rest on one assumption: skills change slowly enough that an inventory stays true longer than it takes to build the training from it. AI moves fastest on exactly the cognitive component the frameworks capture.
Are the fast movers replacing, not retraining?
Let's look at just the last sixteen months.
In June 2025, Amazon's CEO told staff he expects generative AI and agents to reduce the corporate workforce over the next few years, and urged them to "attend workshops and take trainings" (Amazon). Block went from over 10,000 people to just under 6,000 in February 2026, after what it called a strong year. In its shareholder letter, its CEO wrote:
The core thesis is simple. Intelligence tools have changed what it means to build and run a company. We’re already seeing it internally. A significantly smaller team, using the tools we’re building, can do more and do it better. And intelligence tool capabilities are compounding faster every week.
I don’t think we’re early to this realization. I think most companies are late. Within the next year, I believe the majority of companies will reach the same conclusion and make similar structural changes. I’d rather get there honestly and on our own terms than be forced into it reactively.
— Jack Dorsey, Block shareholder letter, Q4 2025 (SEC filing, Exhibit 99.1)
In May, Cloudflare filed a plan to move to an "agentic AI-first operating model" and cut about 20 percent of its workforce (8-K). That was roughly 1,100 people, in a quarter when revenue grew 34 percent, and its CEO called AI and agents "core parts of our workforce" (press release).
Challenger, Gray & Christmas, which tracks announced US layoffs, counts AI as the leading stated reason this year: 120,136 cuts through September, about a fifth of all announced cuts (Challenger, September 2026).
Employers' stated plans in January 2025 were different. The World Economic Forum's survey of over a thousand employers expected 29 of every 100 workers to be upskilled in their current roles and 19 more to be reskilled and redeployed (WEF). A handful of firms does not overturn that survey, but they show where some of the fastest are heading. In my reading, the workforce they are building is AI agents first, with some humans, so an inventory of it has to list agent skills first and human skills as the remainder.
Is most of the industry sticking to its tried and tested approach?
Several figures point away from replacement. Excluding government, Challenger counts announced cuts down 15 percent on the first nine months of 2025, with hiring plans up 3 percent. AI is the leading stated reason for cuts this year, but it was only the fifth most cited in September, and Challenger notes that naming AI can win over investors (Challenger, September 2026). The Census Bureau finds 19.8 percent of US firms using AI as of 3 May 2026, and 37 percent of firms with 250 or more employees (Census Bureau). The Minneapolis Fed, reading Census data, reports that about 96 percent of AI-using firms saw no change in headcount, that 10 percent said AI now does tasks employees used to do, and calls adoption "experimental, task-specific, and uneven" (Minneapolis Fed). Atlassian cut about 10 percent of its team, some 1,600 people, and its CEO wrote: "Our approach is not 'AI replaces people'" (Atlassian).
None of this shows that AI is not displacing people. In my reading, it shows that much of the industry is slowing adoption or avoiding it, and layering AI onto existing structures. Fewer cuts and flat headcount are what that looks like. And a lower cut rate is measured against a base. A company that went from 1,000 people to 600 over three years and then stops cutting shows a quiet year, but 600 may simply be the number it now needs to operate. That example is hypothetical to illustrate that a lower rate of job cuts is natural after a few years of higher job cuts. The Census figures also cover six months of headcount at firms that mostly adopted AI recently. They would show little if replacement starts at the leading firms and spreads outward. That is my reading, and the data does not settle it.
This is where the training-led model still holds: skills are inventoried, gaps are trained, and people use AI as a tool. There, skill-first lives on as a static inventory.
Where a skill inventory breaks
Capability is on a short clock. Model Evaluation and Threat Research (METR) is a nonprofit research group based in Berkeley, California that tests what frontier AI models can do on their own. METR measures how long a task, timed in human hours, an AI model can complete half the time. That horizon has doubled about every seven months since 2019, possibly faster since 2024, on tasks heavily skewed to software and machine learning (arXiv). A March 2026 METR note puts the top model at about 12 hours on the full suite and 7 hours on private tasks, with error bars often a factor of two either way. The same note finds that reasonable alternative modelling choices mostly lower recent horizons, by up to 35 percent, and traces the private-task gap to a handful of RE-Bench tasks (METR). METR measures task completion, not whole jobs, and I use it to read direction, not to size anything.
Forecasts of how many skills change by 2030 run from 39 percent (WEF employers, January 2025, down from 44 percent in 2023) to 70 percent (LinkedIn, the same month, no method disclosed). Both are expectations, not measurements.
A skill inventory has a refresh cycle, and capability has its own. When capability moves faster, an entry describes a task as it used to be done, and training built from it teaches yesterday's method. That is my reasoning, not a published finding.
Exhibit A is a skill inventory entry in miniature. My AI collaborator (Claude) wrote it as a project handover for the AI agent (Claude itself) that would pick up the same work in the next session. Most of it is tool craft, true only of today's tools. One phrase works differently. "Per his preference" names one rule, no clickable nodes, but it also tells the receiving AI to consult its memory of me and apply what it finds to define the output and design accordingly. It is a sophisticated skill: it points to knowledge held elsewhere, states tooling and gives meta-cognitive instruction.
The case for skill-first
METR's tasks skew to software, and many roles sit far from software. And an inventory refreshed continuously by machines might keep up. The real question is whether skill inventories get replaced or only automated.
How long will skill-first last?
In my reading, it is already ending. At the companies moving fastest, replacement has taken the place of training as the response to AI, and the workforce is becoming AI agents with some humans. Elsewhere training continues, but it no longer decides the outcome. Skill-first survives only as an inventory refreshed by machines, and an inventory of that workforce lists agent skills first. A human-maintained inventory grows a little more wrong each quarter.
I have worked with Claude as my execution partner for two years now. In my observation, frontier models in 2024 were Grade 1 students. Today they are graduates with a few years' experience. That is the pace of learning, and it has changed my work. Tasks drop out of my workflows, and my workflows change month on month because I am constantly monitoring AI capabilities and adapting. More importantly, I delegate increasingly complex execution to AI as it demonstrates greater cognitive ability, which lets me focus on creation and architecture, and on reviewing and correcting instead of developing.
My judgement now goes into predicting which emerging skill will matter twelve months out. Most won't, and training for them is waste.
Three questions I use in almost every training situation now:
- How old is the freshest entry in the skills inventory, and what refreshes it?
- Which entries describe tasks an AI system can already perform, and where in the inventory are the skills of the agents that now do them?
- What would one need to see to conclude that skill-first no longer holds?
If one cannot answer the third, skill-first is a belief, not a position.
Editors' Note: The article is conceived and written by the author with Claude employed as research and editing assistant.
Sanjay Mukherjee is the founder of Eduriti, an AI-native product studio, and editor of The Learning Equilibrium. He has spent 34 years across journalism, instructional design and L&D consulting, including 12 years in core operational roles before moving into learning technology. He has served as Vice President, COO and Chief Strategy Advisor for global eLearning companies and currently consults in Fractional Strategic Advisor roles. He is the author of Talking to Claude and other AIs (2025) and is based in Pune, India.